Massive and passive data such as cell phone traces provide samples of the whereabouts and movements of individuals. These are a potential source of information for models of daily activities in a city. The main challenge is that phone traces have low spatial precision and are sparsely sampled in time, which requires a precise set of techniques for mining hidden valuable information they contain. Here we propose a method to reveal activity patterns that emerge from cell phone data by analyzing relational signatures of activity time, duration, and land use. First, we present a method of how to detect stays and extract a robust set of geolocated time stamps that represent trip chains. Second, we show how to cluster activities by combining the detected trip chains with land use data. This is accomplished by modeling the dependencies between activity type, trip scheduling, and land use types via a Relational Markov Network. We apply the method to two different kinds of mobile phone datasets from the metropolitan areas of Vienna, Austria and Boston, USA. The former data includes information from mobility management signals, while the latter are usual Call Detail Records. The resulting trip sequence patterns and activity scheduling from both datasets agree well with their respective city surveys, and we show that the inferred activity clusters are stable across different days and both cities. This method to infer activity patterns from cell phone data allows us to use these as a novel and cheaper data source for activity-based modeling and travel behavior studies.
1 June 2014 in Transportation Research Part C: Emerging Technologies
The problem of map-matching sparse and noisy GPS trajectories to road networks has gained increasing importance in recent years. A common state-of-the-art solution to this problem relies on a Hidden Markov Model (HMM) to identify the most plausible road sequence for a given trajectory. While this approach has been shown to work well on sparse and noisy data, the algorithm has a high computational complexity and becomes slow when working with large trajectories and extended search radii. We propose an optimization to the original approach which significantly reduces the number of state transitions that need to be evaluated in order to identify the correct solution. In experiments with publicly available benchmark data, the proposed optimization yields nearly identical map-matching results as the original algorithm, but reduces the algorithm runtime by up to 45%. We demonstrate that the effects of our optimization become more pronounced when dealing with larger problem spaces and indicate how our approach can be combined with other recent optimizations to further reduce the overall algorithm runtime.
We examine the publicly available OpenCellID database for the purpose of estimating the position of mobile devices along with the positioning accuracy in any given geographic region. We present a method coping with typical distortions and artifacts in the data to compute robust estimates of cell positions and localization error distributions. Applying the method to analyze different regions and mobile networks reveals that the localization errors follow a fat-tailed distribution with similar shapes and properties. We experimentally validate the approach using an independent set of GPS and cell ID data collected by 250 individuals over a period of several weeks. The localization error estimated from OpenCellID data agrees well with the experimental error distribution. While there are significant differences between urban and rural areas, we show that there is only a weak correlation between antenna density and empirical localization errors.